Re: [PATCH v4] HID: New hid-cp2112 driver

[Oliver Neukum <oneukum@suse.de>]

On Thu, 2014-01-30 at 01:52 -0800, Joe Perches wrote:
> On Thu, 2014-01-30 at 08:35 +0100, Oliver Neukum wrote:
> > On Wed, 2014-01-29 at 15:46 -0800, Joe Perches wrote:
> > 
> > > > +static int cp2112_hid_get(struct hid_device *hdev, unsigned char report_number,
> > > > +			  u8 *data, size_t count, unsigned char report_type)
> > > > +{
> > > > +	u8 *buf;
> > > > +	int ret;
> > > > +
> > > > +	buf = kmalloc(count, GFP_KERNEL);
> > > > +	if (!buf)
> > > > +		return -ENOMEM;
> > > > +
> > > > +	ret = hdev->hid_get_raw_report(hdev, report_number, buf, count,
> > > > +				       report_type);
> > > > +	memcpy(data, buf, count);
> > > > +	kfree(buf);
> > > > +	return ret;
> > > 
> > > if the data is going to be copied in data,
> > > why not just use data in hid_get_raw_report
> > > and avoid the malloc?
> > 
> > He must not. It would violate the DMA rules. This function is passed
> > pointers to parts of structures. Access to the structures may race
> > with a DMA to/from the buffer. That violates the DMA rules.
> 
> If this is going to be used to DMA, doesn't
> the alloc need to be GFP_KERNEL | GFP_DMA ?

No. GFP_DMA is needed only for DMA to the ISA bus.

	Regards
		Oliver